How medusa2o/grammar-scoring-engine-on-shl-s-data works
Project type and purpose
• This is a data-science/ML pipeline (not a web app) that ingests raw audio, extracts acoustic features, trains a stack of regression models, and produces a final scoring engine. Its core purpose is to predict human-labeled grammar scores from audio recordings.
Main components
a. Data Ingestion & Storage
• Raw audio files (e.g., WAV/MP3) stored on disk or S3
• Augmented audio store (output of pitch/time augmentations)
b. Preprocessing & Feature Extraction
• Resampling module (to 16 kHz)
• Data augmentation module (pitch shift, time stretch)
• Librosa-based extractor producing 98-dim feature vectors (MFCC, Chroma, Spectral Contrast, Tonnetz, Tempo, RMS, ZCR)
• Intermediate feature store (e.g., CSV, HDF5)
c. Model Training
• Base learners (GBR, RF, ET, SVR, KNN, Ada, HistGB, XGB)
– Each trained in parallel on the same feature set via 5-fold CV
• Out-of-fold prediction collector (stacking dataset)
• Meta-learner (SVR) trained on OOF predictions
• Trained model artifacts (pickles or joblib files)
d. Evaluation & Visualization
• Metrics calculator (RMSE, Pearson r)
• Plot generator (matplotlib/seaborn output.png, output(1).png, etc.)
• Reporting (notebook cells or HTML export)
Relationships & data flow
Raw Audio → Preprocessing → Augmented Audio → Feature Extraction → Features
Features → Base Model Training (parallel) → OOF Predictions → Meta-Model Training → Final Model
Final Model + Test Features → Predictions → Evaluation & Visualization
• CV folds feeding OOF predictions into the stacking layer.
Architectural patterns & design principles
• Modular pipeline pattern: separate stages with clear inputs/outputs
• Ensemble stacking pattern: base layer → meta layer
• Reproducibility: versioned data artifacts and model artifacts
• Scalability: parallel training of base learners
Relevant technologies & libraries
• Python (Jupyter Notebook)
• Librosa (audio preprocessing & feature extraction)
• scikit-learn (GBR, RF, ET, SVR, KNN, AdaBoost, HistGB, pipeline utilities)
• XGBoost (XGB)
• matplotlib/seaborn (visualization)
• joblib or pickle (model persistence)
Essential details to capture
• Data augmentation branching (show original audio → multiple augmented versions)
• Feature dimensionality (98 dims)
• Parallel vs. sequential stages
• Cross-validation feeding into stacking